Playgrounds and Agents: Take Your Prompts Beyond the Chat
Mastering prompting techniques is only step one. Tour the playgrounds of Copilot Studio, Google AI Studio, OpenAI and Claude Workbench, and the four controls you should know how to find.
Mastering prompting techniques is only step one. The real leap comes from taking your prompts into each platform's playground, where you can tune model parameters, compare responses and turn a simple chat into a repeatable, automatable process for a team. We cover Microsoft Copilot / Copilot Studio, Google AI Studio, OpenAI Playground and Claude Workbench.
Microsoft Copilot and Copilot Studio
Copilot has no model of its own: it runs on OpenAI models (GPT-4, GPT-5). Its basic "agents", despite the name, are not agents in the formal AI sense: they are still a chat that cannot take actions on external tools. The same concept elsewhere is a Custom GPT (OpenAI) or a Project (Claude).
When creating one, there are three key elements:
- Instructions: your base prompt. Example: "Act as a data analyst and challenge any insights I give you".
- Knowledge: connect OneDrive or upload files as a reference source, without expanding the context window.
- Tools and starters: toggle capabilities (code writing, image generation) and define conversation starters to make the chat friendlier.
| Uploaded in chat | Uploaded in knowledge | |
|---|---|---|
| Where it lives | Inside the context window | External reference base |
| How it works | Model processes everything at once | Model first searches for relevant passages, extracts them, then uses them in the answer |
| Limit | Bounded by context window | Practically unlimited files |
Copilot Studio is Microsoft's most powerful tool for this purpose. Its agent flows turn a chat into a real AI agent because it can take concrete actions without manual intervention. Example: when a Teams meeting ends, take the transcript, generate a summary and email it to all attendees.
Google AI Studio (the Gemini playground)
Unlike the standard Gemini interface, it lets you:
- Name and save each experiment to remember what you were testing.
- Edit the model's response, since it is part of the context window.
- Build variation trees: start from one prompt, branch with different instructions, and compare results without losing history.
- Use temporary chats: they ignore all prior account information and are deleted when the session closes, giving an isolated environment.
- Run two models side by side, for example Gemini 2.5 Pro (thinks longer) vs. Gemini 2.5 Flash (near-instant). In the example, Flash took 8 s and Pro took 13.5 s.
- See the token count, the unit that measures the context window.
- Set the thinking budget: the number of tokens the model may use to reason before answering. More tokens means deeper reasoning but slower; fewer tokens means faster but shallower.
OpenAI Playground
It works similarly, but billing is per token consumed, independent of any ChatGPT subscription. An approximate cost mentioned in class was about $1 per million tokens (it varies by model, so check current pricing). The same settings are available:
- Temperature.
- Top P.
- Available tools.
- Reasoning effort: minimal, low, medium or high.
Claude Workbench (Anthropic)
It's accessed through Anthropic's developer platform. It includes a prompt generator: you describe the task and the tool writes the prompt for you. This is equivalent to asking any LLM to "act as a prompt engineer", a very useful trick for building better instructions.
- Models: Opus (deep reasoning) and Sonnet (faster responses).
- Settings: temperature, max tokens and thinking budget.
Mini glossary
| Term | What it is |
|---|---|
| Playground | An environment to experiment with a model and adjust its parameters |
| Context window | How much information the model handles at once; everything in the chat counts |
| Knowledge (retrieval) | External documents queried on demand, not loaded entirely into context |
| Token | The unit of measurement for context and billing |
| Temperature | Controls randomness and creativity of responses |
| Top P | Limits the pool of candidate words by cumulative probability |
| Thinking budget / reasoning effort | How much the model reasons before answering |
| AI agent | A system that takes actions on external tools without manual intervention |
Key takeaways
- Interfaces will keep changing; what matters is knowing where to find four controls: temperature, Top P, model tools and reasoning effort.
- Together they give you real control over how the AI responds.
- Files in knowledge don't take up the context window; files in the chat do.
- A chat that only answers is not an agent: an agent executes actions.